Reservoir dam safety monitoring system

Abstract

The invention discloses a reservoir dam safety monitoring system which is mainly composed of a distributed sensing unit, a data fusion unit, a data processing and analyzing unit and a monitoring decision-making unit which are matched with one another. According to the invention, the method has significant advantages in the aspects of data precision, prediction capability, real-time performance, intelligent decision and the like, not only solves the core pain point in the prior art, but also constructs a complete intelligent monitoring system of'perception-fusion-analysis-decision-interaction ', and can effectively overcome the problems in the prior art.

Description

Technical Field

[0001] This invention relates to building safety monitoring technology, specifically to dam safety monitoring technology. Background Technology

[0002] Based on the analysis of the current state of technology in the field of dam safety monitoring, the existing dam safety monitoring technologies mainly have the following technical shortcomings:

[0003] 1. Weak ability to dynamically adapt to environment and operating conditions;

[0004] Existing dam safety monitoring technologies mostly use static or semi-static models, which cannot dynamically adjust monitoring strategies and parameter thresholds according to real-time climate, water level, geological changes, etc., resulting in a significant decrease in monitoring efficiency under complex conditions such as extreme weather and sudden changes in water level.

[0005] 2. Insufficient ability to correlate and fuse data across scales;

[0006] Existing dam safety monitoring technologies are often limited to simple superposition or linear fusion of data at the same scale or from the same source. They lack the ability to deeply correlate and couple data between micro (such as sensor data), meso (such as dam section deformation), and macro (such as reservoir hydrology) data, making it difficult to effectively link and assess local risks with overall stability.

[0007] 3. Limited ability to provide early warning and identify hidden risks;

[0008] Existing dam safety monitoring technologies often rely on historical thresholds or simple statistical models for early warning mechanisms. These mechanisms only provide post-event alarms when data exceeds limits, failing to identify and provide early warnings for hidden risks (such as internal leakage or material fatigue) that do not exceed thresholds but exhibit abnormal causal logic. This results in short warning windows and a high rate of missed reports.

[0009] 4. Poor flexibility in emergency decision-making in sudden and complex scenarios;

[0010] Existing dam safety monitoring technologies lack the ability to quickly simulate and generate dynamic decisions when faced with complex and sudden scenarios (such as "piping + earthquake" or "landslide + sudden rise in water level") that have not been pre-trained for. They rely too much on preset rules and human intervention, resulting in slow response speeds and difficulty in scientifically evaluating the advantages and disadvantages of multiple options.

[0011] 5. Edge perception exhibits low robustness and continuity;

[0012] The edge sensing nodes of existing dam safety monitoring technologies are prone to data drift or failure in harsh environments (high water pressure, strong electromagnetic interference), and lack effective self-diagnosis, data completion and dynamic switching mechanisms for transmission paths, resulting in high data interruption or distortion rates and limited system reliability.

[0013] As can be seen from the above, existing dam safety monitoring technologies can realize the basic safety monitoring process of "data collection-analysis-decision", but they have obvious shortcomings in dynamic environment adaptability, cross-scale data correlation, early warning capability, and decision-making flexibility in emergency scenarios, which greatly affects the accuracy and reliability of the monitoring results of dam safety monitoring technologies. Summary of the Invention

[0014] In view of the shortcomings of existing dam safety monitoring technologies, the present invention aims to provide a reservoir dam safety monitoring technology. This solution can effectively improve the adaptive capability and rapid decision-making and adaptive capability of reservoir dam safety monitoring technology in dynamic and changing environments, as well as improve the reliability and continuity of edge sensing nodes under harsh working conditions, and can effectively overcome the shortcomings of existing technologies.

[0015] To achieve the above objectives, the present invention provides a reservoir dam safety monitoring system, the system comprising:

[0016] The distributed sensing unit includes several sensing modules deployed on the dam body of the reservoir to be monitored, and an edge computing node module that processes the data collected by the several sensing modules. The edge computing node module is configured with an edge robust mechanism to bind the sensor health status with the data validity in real time.

[0017] The data fusion unit connects with the distributed sensing unit to perform hierarchical processing on the received sensing data to form multi-scale data. It also constructs a corresponding scale map structure for each scale data level and then fuses the data using a multi-scale map convolutional fusion model.

[0018] The data processing and analysis unit is connected to the data fusion unit and configured with a spatiotemporal causal inference early warning model and a dynamic spatiotemporal attention prediction model.

[0019] The spatiotemporal causal inference early warning model generates early warning information when the data does not exceed the threshold but the causal logic is abnormal, based on the fused data generated by the data fusion unit and constructing a causal relationship network for dam safety.

[0020] The dynamic spatiotemporal attention prediction model calculates the partition risk weights for the fused data generated by the data fusion unit, adjusts the weights in different time periods, and finally performs mixed prediction calculations based on the weighted features and the identified working conditions.

[0021] The monitoring and decision-making unit interacts with the data processing and analysis unit to construct a corresponding scenario model for the target reservoir dam based on meta-learning. Based on the early warning and prediction data generated by the data processing and analysis unit, it performs twin synchronous simulation, calculates the scenario similarity, fine-tunes the model according to the similarity value, and finally calculates and generates a decision scheme based on the determined scenario model.

[0022] Furthermore, the edge computing node module is configured with a fault diagnosis submodule, a data completion submodule, and a dynamic path switching submodule.

[0023] The fault diagnosis submodule analyzes the data fluctuation characteristics by inputting sensor data, and calculates the sensor module health value based on the analyzed characteristic data and sensor health value.

[0024] The data completion submodule interacts with the fault diagnosis submodule and can complete the missing data of the sensor data by using the inverse distance weighting method when the fault diagnosis submodule detects a fault in the corresponding sensor module.

[0025] The dynamic path switching submodule automatically switches to the backup sensing module based on the preset sensing path for the fault sensing module determined by the fault diagnosis submodule.

[0026] Furthermore, the multi-scale graph convolutional fusion model includes a scale hierarchical module and a scale attention module.

[0027] The scale layering module is used to perform multi-level layering on the transmitted data, generate the corresponding multi-level scale graph structure, and assign initial weights.

[0028] The scale attention module is used to dynamically adjust the weights of each scale by calling the sigmoid function on the multi-level scale graph structure generated by the scale hierarchy module.

[0029] Furthermore, the spatiotemporal causal inference and early warning model is equipped with a causal relationship network module, a real-time causal deviation monitoring module, and a risk tracing and location module.

[0030] The causal relationship network module is configured to use a causal discovery algorithm to uncover strong causal relationships between key parameters of the dam and form a core causal chain.

[0031] The real-time causal deviation monitoring module is configured to interact with the causal relationship network module, calculate the deviation degree for each causal link generated by the causal relationship network module, and monitor the deviation degree of each calculated link. If the deviation degree exceeds the threshold, an early warning is triggered.

[0032] The risk tracing and location module is configured to interact with the real-time causal deviation monitoring module. After the real-time causal deviation monitoring module triggers an early warning, it can locate the source of risk for causal links with deviations exceeding a threshold using a causal reverse reasoning algorithm.

[0033] Furthermore, the dynamic spatiotemporal attention prediction model is configured with a spatial dynamic attention module, a temporal dynamic attention module, and a hybrid prediction output module.

[0034] The spatial dynamic attention module is configured to divide the dam into N monitoring units and dynamically adjust the attention weight by calculating the risk contribution of each unit.

[0035] The time-based dynamic attention module is configured to divide the time series into sensitive time periods, assign sensitivity coefficients to each of the divided sensitive time periods, and adjust the weights accordingly through the time period sensitivity function.

[0036] The hybrid prediction output module introduces real-time operating data, selects the optimal prediction sub-model based on the real-time operating conditions, and generates deformation prediction values.

[0037] Furthermore, in the meta-training phase, the meta-reinforcement learning dynamic decision engine constructs a database of dam emergency scenarios and uses the MAML algorithm to train the basic decision model.

[0038] Furthermore, in the real-time decision-making stage, the meta-reinforcement learning dynamic decision engine first performs digital twin synchronization, inputting real-time collected sensor data and external environment data into the dam's digital twin model to simulate the dam's stress state under the current scenario; then, it performs scene matching and fine-tuning, comparing the similarity between the current scene and the scene library. If the similarity is less than the corresponding threshold, it calls the meta-learning algorithm to quickly fine-tune the decision model based on the virtual samples simulated by the digital twin; finally, it outputs multi-objective decisions, generating decision schemes for new scenarios.

[0039] The solution provided by this invention has the following advantages over the prior art:

[0040] (1) This system solution can effectively improve the adaptive capability in a dynamic and ever-changing environment and realize the real-time optimization and adjustment of monitoring strategies, model parameters and decision thresholds;

[0041] (2) This system solution establishes an efficient correlation and comprehensive analysis mechanism for "micro-meta-macro" data, effectively breaking through the technical bottleneck of multi-scale and multi-modal data fusion;

[0042] (3) This system solution can realize the transformation from "post-event alarm" to "pre-event warning", significantly extending the warning window and improving the ability to identify hidden and gradual risks in the early stage;

[0043] (4) This system solution effectively enhances the ability to make rapid decisions and adapt in unknown emergency scenarios, reduces reliance on human experience, and improves the speed and scientific nature of emergency response.

[0044] (5) This system solution can improve the reliability and continuity of edge sensing nodes under harsh working conditions, and build a robust sensing system with self-diagnosis of faults, self-completion of data and self-switching of transmission. Attached Figure Description

[0045] The present invention will be further described below with reference to the accompanying drawings and specific embodiments.

[0046] Figure 1 This is a schematic diagram illustrating the structural principle of the reservoir dam safety monitoring system in this invention.

[0047] Figure 2 This is an example diagram illustrating the structure of a reservoir dam safety monitoring system in this invention.

[0048] Figure 3 This is a flowchart of the reservoir dam safety monitoring process in an example of the present invention;

[0049] Figure 4 This is a schematic diagram illustrating the deployment of the distributed wireless communication module in an example of the present invention;

[0050] Figure 5 This is a schematic diagram of the data processing principle of the multi-source data fusion module in this invention. Detailed Implementation

[0051] To make the technical means, creative features, objectives and effects of this invention easier to understand, the invention will be further described below with reference to specific illustrations.

[0052] Combination Figure 1 As shown, the reservoir dam safety monitoring system 100 provided by the present invention mainly consists of four parts that cooperate with each other: a distributed sensing unit 110, a data fusion unit 120, a data processing and analysis unit 130, and a monitoring and decision-making unit 140.

[0053] The distributed sensing unit 110 constitutes the sensing layer of the entire system. It mainly consists of two parts: a number of sensing modules deployed on the dam body of the reservoir to be monitored, and an edge computing node module that processes the data collected by the sensing modules.

[0054] Several sensing modules are deployed in various ways at synchronous locations on the reservoir dam body to collect various status data and / or surrounding environmental data from different parts of the dam body in real time.

[0055] There are no restrictions on the number, type, deployment method, or type of data collected from the sensor modules; these can be determined based on actual needs.

[0056] The edge computing node module is configured with an edge robust mechanism to bind the sensor health status with the validity of the data in real time.

[0057] Specifically, the edge computing node module, based on the configured edge robust mechanism, first calculates and evaluates the health value of the corresponding sensor in the sensing module based on the data collected by the sensing module, and calls the real-time data and historical data of the adjacent sensor for the faulty sensor, and uses the IDW algorithm to complete it; at the same time, it triggers the switching of the sensing data acquisition path.

[0058] This distributed sensing unit 110 addresses the problems of "failure leading to data interruption" and "drift leading to misjudgment" in existing edge sensors. Through a three-in-one edge robust mechanism of "fault diagnosis - data completion - path switching", it effectively binds the sensor health status with data validity in real time, thereby solving the problem of continuous data acquisition in harsh environments (such as high water pressure and strong electromagnetic interference).

[0059] The data fusion unit 120 constitutes the data fusion layer of the entire system. This data fusion unit 120 can connect with the distributed sensing unit 110 and is configured with a multi-scale graph convolutional fusion model of "hierarchical weighting-scale interaction". For the data collected and uploaded by the distributed sensing unit 110, the data is divided and processed based on multiple scales to form multi-level scale data. For each level of scale data, a corresponding scale graph structure is constructed. Then, the data is fused through the multi-scale graph convolutional fusion model. Finally, the corresponding weights are dynamically adjusted based on the scale attention module for the fused data.

[0060] This data fusion unit 120 addresses the problem in existing technologies where GNNs can only process "data at the same scale" (such as data that are all sensor-level data) and cannot correlate "micro-meso-macro" cross-scale information (such as micro crack data → meso dam section deformation → macro reservoir hydrology). Specifically, it adopts a multi-scale fusion model of "hierarchical weighting and scale interaction" to effectively realize the correlation analysis between local damage and overall stability of the dam.

[0061] The data processing and analysis unit 130 constitutes the analysis layer of the entire system. The data processing and analysis unit 130 is connected to the data fusion unit 120. At the same time, the data processing and analysis unit 130 is equipped with a spatiotemporal causal inference early warning model and a dynamic spatiotemporal attention prediction model. Through these two models, the data after being fused by the data fusion unit 120 is analyzed from two directions.

[0062] First, the spatiotemporal causal inference early warning model constructs a "causal relationship network" for dam safety, providing early warnings when data does not exceed the threshold but "causal logic is abnormal".

[0063] Specifically, this spatiotemporal causal inference early warning model uses a PC algorithm to mine causal chains from the fused data generated by the data fusion unit 120, fits the expected value accordingly, calculates the deviation value, and generates an early warning when the deviation value exceeds the threshold.

[0064] This spatiotemporal causal inference early warning model is based on an innovative "causal chain mining-deviation early warning" algorithm. By constructing a "causal relationship network" for dam safety, it can provide early warnings when the data does not exceed the threshold but the "causal logic is abnormal". This can advance the warning time from "1 hour after the event" to "12-24 hours before the event", thus effectively overcoming the problem that existing solutions rely on "data exceeding the threshold alarm" (post-event response) and cannot achieve "risk early warning".

[0065] Furthermore, the dynamic spatiotemporal attention prediction model is equipped with a "spatiotemporal dual-dimensional dynamic attention mechanism," which allows the model to focus on "high-impact areas + high-sensitivity periods" in real time, thereby improving the prediction accuracy under complex working conditions.

[0066] Specifically, this dynamic spatiotemporal attention prediction model calculates the partition risk weights for the fused data generated by the data fusion unit 120, adjusts the weights in different time periods, and finally performs mixed prediction calculations based on the weighted features and the identified working conditions.

[0067] This dynamic spatiotemporal attention prediction model is based on an innovative spatiotemporal dual-dimensional dynamic attention mechanism, which can effectively overcome the problem that the existing LSTM-Transformer model, which uses "static attention weights", cannot adapt to the "spatiotemporal dynamic changes" of the dam (such as the differences in seepage effects in different areas of the dam during the rainy season and the changes in the dam surface temperature gradient in winter).

[0068] The monitoring and decision-making unit 140 constitutes the decision-making layer of the entire system. It interacts with the data processing and analysis unit 130 and is equipped with a meta-reinforcement learning dynamic decision-making engine. Through the dynamic decision-making mechanism of "meta-learning + digital twin", it achieves the decision-making capability of "quickly adapting a small number of samples to new scenarios".

[0069] Specifically, the monitoring and decision-making unit 140 constructs a corresponding scenario model for the target reservoir dam based on meta-learning. On this basis, based on the early warning and prediction data generated by the data processing and analysis unit 130, it performs twin synchronous simulation, calculates the scenario similarity, fine-tunes the model according to the similarity value, and finally calculates and generates a decision scheme based on the determined scenario model.

[0070] This monitoring and decision-making unit 140 can effectively overcome the problem that existing solutions using reinforcement learning can only cope with "pre-trained scenarios" (such as preset "seepage mutation" and "stress over-limit"), but fail when faced with unpre-trained sudden scenarios (such as "local piping in the dam body + earthquake aftershocks" and "landslide in the reservoir area + sudden rise in water level").

[0071] Regarding the reservoir dam safety monitoring system 100 provided by this invention, the following details the specific implementation schemes of each functional unit in the system and the equipment components that may be involved.

[0072] In the specific implementation of the distributed sensing unit 110 in this system, the sensing module preferably adopts a distributed high-precision sensor array. For example, a three-dimensional seepage detection sensor array can be evenly distributed at different depths of the dam, and a fiber Bragg grating strain sensor can be used, which has strong anti-electromagnetic interference capability and high measurement accuracy.

[0073] Furthermore, the edge computing node module in the distributed sensing unit 110 is configured with a fault diagnosis submodule, a data completion submodule, and a dynamic path switching submodule. The edge computing processing of the sensor-collected data is realized through the cooperation between the fault diagnosis submodule, the data completion submodule, and the dynamic path switching submodule.

[0074] Specifically, the fault diagnosis submodule is mainly composed of a data fluctuation characteristic analysis module and a sensor health assessment module.

[0075] The data fluctuation feature analysis module analyzes the data fluctuation characteristics of the input sensor data, such as the frequency of sudden jumps and the mean drift, and transmits the analyzed feature data to the sensor health assessment module. The sensor health assessment module calculates the sensor health value based on this and outputs the sensor module health value in real time.

[0076] As further explanation, this fault diagnosis submodule can be implemented based on the lightweight MobileNetV3 model, that is, the data fluctuation feature analysis module and the sensor health assessment module are integrated on the lightweight MobileNetV3 model.

[0077] As further explanation, the sensor health assessment module here can be implemented using the corresponding sensor health assessment formula.

[0078] This solution quantifies the health status of edge sensors (such as fiber Bragg grating strain sensors and pressure sensors) by integrating two core indicators: data jump frequency and mean drift. The corresponding sensor health assessment formula is as follows:

[0079]

[0080] Variable definition:

[0081] H: Sensor health status (value range: 0-100, H≥80 is normal, 50≤H<80 is slight drift, H<50 is severe failure);

[0082] ω1, ω2: Weighting coefficients ω1 = 0.4, ω2 = 0.6, with priority given to mean drift;

[0083] f: Current data jump frequency (unit: times / hour, jump is defined as a single data fluctuation exceeding twice the normal range);

[0084] f max Maximum allowable jump frequency for the sensor (empirical value: pressure sensor = 2 times / hour, strain sensor = 1 time / hour);

[0085] The current 1-hour average data (e.g., pressure unit: kPa, strain unit: με);

[0086] Average value of sensor under normal operating conditions (taken from data storage module with no abnormal data for the past 30 days); Δx max : Maximum allowable mean drift of the sensor (pressure sensor = 5 kPa, strain sensor = 20 με).

[0087] As a further example, when this fault diagnosis submodule assesses the health of the corresponding sensors in the sensing module based on the sensor health assessment module, it analyzes data fluctuation characteristics, such as the "sudden jump frequency" and "mean drift" of the pressure sensor, and outputs the sensor health value from 0 to 100 points in real time.

[0088] If the health value is ≥80: the data is used directly;

[0089] If 50 ≤ health value < 80: it is judged as "mild drift", and data calibration is initiated (linear correction based on the normal data of the past 1 hour);

[0090] If the health value is less than 50, it is judged as a "severe failure" and the backup plan is triggered.

[0091] The data completion submodule in the edge computing node module interacts with the fault diagnosis submodule. When the fault diagnosis submodule detects a fault in the corresponding sensor module, it can use the "spatiotemporal interpolation completion algorithm" to call the real-time data of three adjacent sensors of the same type plus historical data of the same period to construct a three-dimensional spatiotemporal matrix and use the inverse distance weighting method (IDW) to complete the missing data.

[0092] Specifically, this data completion submodule is mainly composed of an adjacent data calling unit, a historical data retrieval unit, a three-dimensional spatiotemporal matrix module, and an IDW calculation unit.

[0093] Among them, the adjacent data retrieval unit is used to extract real-time data and historical data collected by adjacent sensors of the fault sensor after receiving a fault signal.

[0094] The historical data retrieval unit interacts with the adjacent data retrieval unit to extract real-time and historical data of the three adjacent sensors of the faulty sensor from the data extracted by the adjacent data retrieval unit.

[0095] The three-dimensional spatiotemporal matrix module interacts with the historical data retrieval unit to construct a corresponding three-dimensional matrix based on the real-time data of three adjacent sensors extracted by the historical data retrieval unit and the historical data.

[0096] The IDW computing unit interacts with the 3D spatiotemporal matrix module, and calls the IDW algorithm to perform data completion calculations based on the 3D matrix constructed by the 3D spatiotemporal matrix module.

[0097] As further explanation, the inverse distance weighted method (IDW) in the data completion submodule here can be implemented using the following calculation formula:

[0098]

[0099] Variable definition:

[0100] x 补 Faulty sensor data completion;

[0101] x i Real-time data of the i-th adjacent normal sensor (i = 1, 2, 3, adjacent is defined as spatial distance < 10m);

[0102] d i : Spatial distance between the faulty sensor and the i-th adjacent sensor (unit: m, calculated based on the coordinates of the three-dimensional seepage sensor array);

[0103] x h : Normal data of faulty sensors from the same historical period (e.g., the same time yesterday) (taken from data storage module 19);

[0104] d h Time distance weight (fixed value = 1, dimensionless);

[0105] α: Historical data confidence coefficient (α = 0.3 for normal operating conditions, α = 0.1 for extreme weather, reducing the weight of historical data).

[0106] Thus, the inverse distance weighted method (IDW) formed by the above formula can fill in missing values ​​based on real-time data and historical data from three adjacent sensors of the same type when a sensor fails.

[0107] The dynamic path switching submodule in the edge computing node module interacts with the fault diagnosis submodule and the data completion submodule. It can automatically switch to the backup sensor based on the preset "primary-backup-emergency" three-level sensing path for the faulty sensor determined by the fault diagnosis submodule.

[0108] It should be noted that the preferred preset sensing paths in this dynamic path switching submodule are as follows:

[0109] The main path uses high-precision sensors in the critical area, the backup path uses the same type of sensors next to the main path, and the emergency path uses low-power sensors for the entire dam.

[0110] As further explanation, the dynamic path switching submodule and the data completion submodule are configured to complement each other. The data completion submodule operates during the transition period when the dynamic path switching submodule switches the sensor path. That is, after the fault diagnosis submodule diagnoses a faulty sensor, during the transition period when the dynamic path switching submodule switches the sensor path, the data completion submodule completes the data for the faulty sensor. After the switch is completed, the new sensor data is used.

[0111] As an example, when the seepage pressure sensor inside the dam body fails, this dynamic path switching submodule can automatically control the switch to the backup sensor in the shallow layer of the dam foundation, with a switching delay of less than 50ms, ensuring that the data is not interrupted.

[0112] The distributed sensing unit 110 formed by the above scheme is based on the three-in-one edge robust mechanism of "fault diagnosis-data completion-path switching". Compared with the existing edge model scheme, the data acquisition continuity is improved to 99.8% (original 85%), the false alarm rate caused by sensor failure is reduced by 60%, and it can maintain 72 hours of effective data output without manual on-site sensor replacement.

[0113] The multi-scale graph convolution fusion model in the data processing and analysis unit 120 of this system mainly includes two parts: a scale layering module and a scale attention module.

[0114] The scale layering module is used to perform multi-level layering on the transmitted data and generate the corresponding multi-level scale graph structure and initial weights.

[0115] This scale-based hierarchical module is mainly composed of scale division units, graph structure construction units, and weight initialization units.

[0116] The scale division unit classifies the transmitted data into multiple levels based on the scale, such as dividing the data into three levels: micro, meso, and macro.

[0117] The graph structure building unit is used to build a graph structure for each level of data distribution divided by the scale division unit. In the constructed graph structure, the nodes are the data sources and the edges are the spatiotemporal relationships.

[0118] The weight initialization unit is used to initialize the weights of the graph structure constructed by the graph structure construction unit, such as 40% for micro, 35% for meso, and 25% for macro.

[0119] Accordingly, this multi-scale graph convolutional fusion model can generate three-level scale graphs and initial weights, which fit the dam scenario, and the graph structure reflects the data correlation.

[0120] The scale attention module is used to dynamically adjust the weights of each scale by calling the sigmoid function on the multi-level scale graph structure generated by the scale hierarchy module.

[0121] The attention module at this scale is mainly composed of a condition recognition unit, a sigmoid weight calculation unit, and a weight update unit.

[0122] The working condition identification unit is used to identify the on-site working conditions and calculate the corresponding scale sensitivity coefficient based on the identified working conditions.

[0123] The igmoid weight calculation unit calculates the weights corresponding to the multi-level scale map structure based on the scale sensitivity coefficient determined by the working condition identification unit.

[0124] The weight update unit dynamically updates the weights of the multi-scale graph structure based on the weights calculated by the igmoid weight calculation unit.

[0125] To further explain, this multi-scale graph convolutional fusion model employs multi-scale GCN convolution to achieve scale layering, and incorporates scale weights into the graph convolutional layers to enable cross-scale information transfer. The specific formula is as follows:

[0126]

[0127] X′ (l+1) The fused feature matrix output by the (l+1)th convolutional layer;

[0128] Normalized graph adjacency matrix (reflecting the spatial association between sensors / dam sections / hydrological stations);

[0129] The input feature matrix of the l-th layer at the m-th scale (e.g., strain sensor data for the micro-layer and water level data for the macro-layer);

[0130] The convolutional weight matrix of the l-th layer at the m-th scale (model training and learning parameters);

[0131] b (l) : The bias vector of the l-th layer;

[0132] σ: Activation function, preferably ReLU function, σ(x)=max(0,x).

[0133] Based on this, the scale weights are dynamically adjusted using the following formula:

[0134] The fusion weights for micro / meso / macro scales are dynamically assigned based on real-time operating conditions (rainfall, water level), as shown in the following formula:

[0135]

[0136] Variable definition:

[0137] ω m : The weight of the m-th scale (m∈{micro, meso, macro}, ∑ω m =1);

[0138] λ m Scale sensitivity coefficient (microscopic λ) 微 =1.2, mesoscopic λ 中 =1.0, macroscopic λ 宏 =1.5);

[0139] S m Trigger value for standard operating conditions (dimensionless, value 0-1):

[0140] Macroscale S 宏 : (When the water level rises by more than 1 meter in a single day, S) 宏 =1);

[0141] Microscale S 微 : (When the crack extends more than 0.2 mm / day) 微 =1);

[0142] Mesoscale S 中 :S 中 =1-max(S) 宏 ,S 微 (Balancing meso-level weights under non-extreme operating conditions).

[0143] As an example: when the daily rise in water level in the reservoir exceeds 1m, the weight of the macro scale increases from 30% to 50%, while the weight of the micro scale decreases from 40% to 25%, prioritizing the focus on "the impact of a sudden rise in water level on the overall stress of the dam".

[0144] When the microscopic sensor detects a crack propagation rate > 0.2 mm / day, the weight of the microscopic scale is increased to 55%, focusing on the correlation between "local cracks and the stress distribution of the dam section".

[0145] As a further example, the multi-scale graph convolutional fusion model formed based on the above scheme can divide the data transmitted by the distributed sensing unit into three scales and construct a "scale graph structure":

[0146]

[0147] In practical applications, this can increase the accuracy of comprehensive assessment by cross-scale data fusion from the original 75% to 92%, effectively realizing the correlation prediction of "local cracks → dam section instability risk" and avoiding misjudgments of "only looking at the micro level and missing the macro level".

[0148] The spatiotemporal causal inference and early warning model in the data processing and analysis unit 130 of this system is equipped with a causal relationship network module, a real-time causal deviation monitoring module, and a risk tracing and positioning module. The corresponding functions are achieved through the cooperation between the causal relationship network module, the real-time causal deviation monitoring module, and the risk tracing and positioning module.

[0149] The causal relationship network module is designed to use a causal discovery algorithm (PC algorithm + causal graph learning) to uncover strong causal relationships between key parameters of the dam and form a core causal chain.

[0150] Specifically, this causal relationship network module is composed of a data preprocessing unit, a PC algorithm engine unit, a causal graph generation unit, and an expected value calculation unit.

[0151] The data preprocessing unit is used to clean the received fused data.

[0152] The PC algorithm engine unit is used to mine strong causal chains from cleaned and fused data using PC algorithms.

[0153] The causal graph generation unit constructs a causal graph based on the causal chains generated by the PC algorithm engine unit;

[0154] The expected value calculation unit calculates the causal expected value by fitting the expected value formula to the causal graph constructed by the causal graph generation unit.

[0155] As an example, taking the causal chain of "rainfall → water level" as an example, the causal relationship is fitted based on historical data, and the formula is as follows:

[0156] y 预 =k·x+c+ε

[0157] Variable definition:

[0158] y预 : Expected water level (unit: m);

[0159] x: Rainfall (unit: mm / 24h);

[0160] k, c: Causality coefficients (obtained by linear regression of rainfall-water level data for the same period in the past 3 years, e.g., k = 0.012 m / mm, c = 12.5 m);

[0161] ε: Error correction term (extreme weather ε=+0.3m, drought period ε=-0.2m).

[0162] As a further example, this solution, based on the causal relationship network module, can form the following causal chain:

[0163] Basin rainfall → reservoir water level → dam seepage pressure → dam foundation rock stress → dam displacement → risk of crack propagation.

[0164] Each causal link is labeled with a "normal causal threshold". For example, for every 100mm increase in rainfall, the water level should rise by 0.8-1.2m. If it exceeds this range, it is judged as "causal deviation".

[0165] The real-time causal deviation monitoring module in this spatiotemporal causal inference early warning model is configured to interact with the causal relationship network module. It can calculate the deviation degree of each causal link generated by the causal relationship network module, monitor the deviation degree of each calculated link, and trigger an early warning when the deviation degree exceeds the threshold.

[0166] As further explanation, this real-time causal deviation monitoring module calculates the deviation of each causal link by dividing the difference between the current actual value and the expected causal value by the expected causal value. The corresponding calculation formula is as follows:

[0167]

[0168] Variable definition:

[0169] D: Causality deviation (D>15% and triggers an alert if it lasts for 5 minutes);

[0170] y 实 Real-time water level monitoring value (taken from the water level sensor in multi-parameter detection module 2);

[0171] y 预 : Expected causal value of water level.

[0172] Based on this, when monitoring each calculated link deviation, if any link deviation is greater than 15% and lasts for 5 minutes, an early warning is triggered. This quantifies the degree of deviation between the actual data and the expected causal value, triggering an early warning.

[0173] As an example, when rainfall increases by 100mm, the water level rises by only 0.5m (deviation -41.7%), which may indicate that "there is a hidden seepage channel inside the dam." At this time, the seepage pressure data has not exceeded the threshold, but the causal deviation has exposed the risk, and an early warning is issued 18 hours in advance.

[0174] The risk tracing and location module in this spatiotemporal causal inference early warning model is configured to interact with the real-time causal deviation monitoring module. After the real-time causal deviation monitoring module triggers an early warning, it can locate the source of risk for causal links with deviations exceeding the threshold through a "causal reverse inference algorithm".

[0175] As further explanation, when locating the source of risk based on the causal inverse reasoning algorithm, this risk tracing and location module follows the inverse reasoning process as follows:

[0176] Connect the deviation link → sort by deviation degree → trace upstream parameters in reverse → locate the risk source in combination with dam structure.

[0177] As an example, if there is a deviation in the "seepage pressure → stress" link, the dam foundation seepage prevention curtain should be checked first; if there is a deviation in the "water level → seepage pressure" link, the dam body concrete density should be checked first.

[0178] Based on the aforementioned spatiotemporal causal inference early warning model, this solution can advance the risk warning time to 12-24 hours (previously <1 hour), increase the detection rate of hidden risks (such as internal leakage channels) by 75%, and avoid missed judgments when "the data has not exceeded the threshold but the risk has already occurred".

[0179] In this system, the dynamic spatiotemporal attention prediction model in the data processing and analysis unit 130 is implemented by configuring a spatial dynamic attention module, a temporal dynamic attention module, and a hybrid prediction output module, and the spatial dynamic attention module, the temporal dynamic attention module, and the hybrid prediction output module work together to achieve the corresponding functions.

[0180] The spatial dynamic attention module is configured to divide the dam into N monitoring units and dynamically adjust the attention weight by calculating the risk contribution of each unit.

[0181] As further explanation, when calculating the risk contribution of each unit, the risk contribution of each unit is specifically determined by calculating the historical failure rate of the sensors within the unit.

[0182] As further explanation, this spatial dynamic attention module allocates spatial attention based on the historical failure rate of the sensor, and specifically uses the following spatial attention weight formula to dynamically adjust the attention weight.

[0183] The formula is as follows:

[0184]

[0185] Variable definition:

[0186] a spatial (i): Spatial attention weights of the i-th sensor;

[0187] r i Historical failure rate of the i-th sensor ( Retrieved from data storage module 19);

[0188] γ: Attention amplification factor (γ = 2.0, enhancing the attention of sensors with high failure rates);

[0189] N: Total number of sensors of the same type (e.g., the number of sensors in the three-dimensional seepage sensor array 5).

[0190] As an example, this spatial dynamic attention module divides the dam into N "monitoring units" (e.g., a 50m × 50m grid), calculates the risk contribution of each unit based on weighted statistics of historical fault data, and dynamically adjusts the attention weights accordingly.

[0191] The attention weight of the dam heel region (historical failure rate 35%) is 7 times that of the dam crest region (historical failure rate 5%), and the model prioritizes learning the deformation patterns of high-risk regions.

[0192] The time dynamic attention module in this dynamic spatiotemporal attention prediction model is configured to divide the time series into sensitive periods, assign sensitivity coefficients to each of the divided sensitive periods, and adjust the weights accordingly through the "period sensitivity function".

[0193] It should be noted that the temporal dynamic attention module is preferably configured to calculate weights in parallel with the spatial dynamic attention module, and then the weights are superimposed to obtain the joint weights, thereby improving prediction accuracy.

[0194] As an example, the non-sensitive period here is a 24-hour time attention window; the sensitive period is an extended 72-hour time attention window, with the weight of the most recent 12 hours' data increased to 60%, focusing on capturing short-term sudden trends.

[0195] As further explanation, this time-based dynamic attention module uses a time-based attention weight formula to adjust weights, as follows:

[0196]

[0197] Variable definition:

[0198] a temporal (t): Temporal attention weight at time t;

[0199] s t Sensitivity of time period at time t (flood season / earthquake season s) t =1.0, daily s t =0.3);

[0200] β: Time weighting coefficient β = 1.5);

[0201] T: Prediction window duration (in the file, if the prediction window is ≥72 hours, T=72).

[0202] The hybrid prediction output module in this dynamic spatiotemporal attention prediction model is configured to be based on LSTM-Transformer, and by introducing real-time operating data, select the optimal prediction sub-model according to the real-time operating conditions to generate deformation prediction values.

[0203] As an example, taking dam displacement as an example, this hybrid prediction output module integrates spatiotemporal attention and determines the final predicted value through the following formula:

[0204]

[0205] Variable definition:

[0206] Predicted dam displacement (unit: mm, prediction error ≤ 0.2 mm);

[0207] d i,t : Historical data of the i-th displacement sensor at time t;

[0208] W, b: Model training parameters (learned by the LSTM-Transformer hybrid architecture).

[0209] Based on the dynamic spatiotemporal attention prediction model formed above, this solution can reduce the dam displacement prediction error from ≤0.5mm to ≤0.2mm, and improve the prediction accuracy by 40% under extreme conditions (such as a once-in-a-century flood), effectively solving the core problem of "adapting static models to dynamic conditions".

[0210] When the meta-reinforcement learning dynamic decision engine in the monitoring decision unit 140 of this system is applied, it mainly includes two stages: the meta-training stage and the real-time decision stage.

[0211] First, in the meta-training phase, a database of dam emergency scenarios is built based on meta-learning, and the basic decision-making model is trained using the MAML algorithm, enabling the model to have the ability to quickly fine-tune.

[0212] Specifically, first, a library of multiple types of scenarios is built, then the DQN model is initialized, and then the MAML algorithm is used to train multiple scenarios alternately, with the inner loop fine-tuning and the outer loop iterating over the initial parameters, thereby forming the target scenario model.

[0213] Furthermore, in the real-time decision-making phase, the first step is to synchronize the digital twin by inputting the real-time sensor data and external environmental data into the dam's digital twin model to simulate the stress state of the dam body under the current scenario.

[0214] Next, scene matching and fine-tuning are performed. The similarity between the current scene and the scene library is compared. If the similarity is less than the corresponding threshold, the meta-learning algorithm is called to quickly fine-tune the decision model based on the virtual samples simulated by the digital twin. Finally, multi-objective decision output is performed to generate a "priority-cost" two-dimensional decision scheme for the new scene.

[0215] As further explanation, this solution determines the similarity between the current scene and the historical scene library using the following formula:

[0216]

[0217] Variable definition:

[0218] Sim: Scene similarity (Sim≥60% indicates a similar scene, Sim<60% indicates a new scene);

[0219] Current scene feature vector (dimension = 5, including: seepage pressure anomaly, strain exceeding threshold, water level rise, rainfall, and earthquake intensity);

[0220] The feature vector of a scene in the historical scene library (taken from the 23 types of basic scene data stored in modular functional component 16).

[0221] When generating "priority-cost" dual-dimensional decision-making solutions for new scenarios, this solution introduces a decision benefit function formula to evaluate the overall benefit of the decision-making solution (risk reduction rate - implementation cost), as follows:

[0222] R = α·R risk -(1-α)-C cost ;

[0223] Variable definition:

[0224] R: The benefit value of the decision-making scheme (the larger the R, the better the scheme, and the scheme with the largest R is selected first);

[0225] α: Return weight (α = 0.7, prioritizing risk reduction);

[0226] R risk Risk reduction rate of the solution (e.g., "partial sealing of piping") R risk =85%);

[0227] C costImplementation cost of the scheme (normalized to 0-1, such as "small flow flood discharge" C) cost =0.3, "Large-flow flood discharge" C cost =0.8).

[0228] As an example, this monitoring decision unit 140, based on a meta-reinforcement learning dynamic decision engine, generates the following "priority-cost" two-dimensional decision scheme for a new scenario:

[0229] Decision-making scheme Emergency measures Risk reduction rate Implementation costs Priority Option 1 Localized sealing of piping and small-flow flood discharge 85％ Low 1 Option 2 Dam reinforcement + high-flow-rate flood discharge 98％ high 2

[0230] Based on the monitoring and decision-making unit formed by the above scheme, this solution can reduce the decision adaptation time for new emergency scenarios from "4 hours of manual intervention" to "5 minutes of automatic adaptation", improve the effectiveness of the decision-making scheme implementation by 80%, and effectively avoid the risk of "no decision available outside the pre-training scenario".

[0231] The following specific examples further illustrate the reservoir dam safety monitoring system solution proposed in this invention.

[0232] See Figure 2 The diagram shows the configuration of a reservoir dam safety monitoring system based on the present invention.

[0233] Based on the diagram, the reservoir dam safety monitoring system mainly includes a data acquisition unit 1, a multi-parameter detection module 2, a high-precision sensing unit 3, a data fusion unit 9, a multi-source data fusion module 10, a data processing and analysis unit 11, a visualization interactive interface unit 12, a distributed wireless communication module 15, modular functional components 16, an alarm module 17, a remote control module 18, and a data storage module 19.

[0234] Among them, the data acquisition unit 1, together with the multi-parameter detection module 2 and the high-precision sensing unit 3, forms a corresponding distributed sensing unit.

[0235] The multi-parameter detection module 2 and the high-precision sensing unit 3 are deployed together on the dam body of the reservoir to be monitored, and are used to collect various status data and / or surrounding environmental data of different parts of the dam body in real time.

[0236] The multi-parameter detection module 2 mainly includes a temperature sensor 4, a three-dimensional seepage detection sensor array 5, and a stress-strain sensor 6.

[0237] Such a multi-parameter detection module (seepage, stress, temperature) collects data simultaneously, covering multiple dimensions of data from "dam body to dam foundation to external environment", providing a complete data foundation for subsequent analysis.

[0238] Among them, the three-dimensional seepage detection sensor array is evenly distributed at different depths and locations of the dam, which can capture the three-dimensional distribution of the seepage field inside the dam body and avoid "local data loss".

[0239] The high-precision sensing unit 3 includes a pressure sensor 7 and a strain sensor 8 based on a fiber Bragg grating.

[0240] The fiber Bragg grating-based strain sensor (with electromagnetic interference resistance >60dB and measurement accuracy ±0.1%) can still stably output high-precision strain data in harsh electromagnetic and underwater environments.

[0241] The data acquisition unit 1 is built based on the edge computing node module in this scheme. It can interact with the multi-parameter detection module 2 and the high-precision sensing unit 3 to complete the edge preprocessing of the sensing data.

[0242] The data fusion unit 9 interacts with the data acquisition unit 1 to form the data fusion layer of the entire system, which is used to complete the multi-source data fusion processing.

[0243] The data fusion unit 9 is equipped with a corresponding multi-source data fusion module 10. The multi-source data fusion module 10 is built based on graph neural network (GNN). It defines sensor data, meteorological and hydrological data (rainfall, water level, etc.), and geological models as a "node-edge" graph structure. Through cross-modal information transmission, it explores the dynamic relationship between the dam body and the external environment. The comprehensive evaluation accuracy is 25% higher than that of the existing technology, breaking through the limitations of linear fusion.

[0244] The data processing and analysis unit 11 constitutes the data analysis layer of the entire system. It combines real-time hydrological data (the water level in the reservoir is stable at 145m with no sudden rise) and geological data (the abnormal area is a sandy gravel stratum with a large permeability coefficient) from the multi-source data fusion module 10, and conducts risk assessment through a spatiotemporal causal inference algorithm.

[0245] The visual interactive interface unit 12 integrates corresponding virtual reality unit 13 and augmented reality unit 14, which can intuitively display the dam's safety status through charts, animations, and real-scene overlays (such as AR guidance for locating and repairing seepage points), and provide data query, analysis, and simulation prediction functions to improve the efficiency of operation and maintenance personnel's perception of the safety status and decision-making.

[0246] The distributed wireless communication module 15 establishes data communication with the data processing and analysis unit 11, the visualization and interactive interface unit 12, the alarm module 17, and the remote control module 18. Specifically, it adopts a 5G / LoRa distributed wireless communication architecture, combined with wireless relay nodes to expand signal coverage, and combines edge computing to transmit only key feature data. The communication latency is <200ms, and the entire process from data acquisition to policy generation takes <1s. The operation and maintenance cost is reduced by 30%, which significantly improves real-time performance and economy compared with existing technologies.

[0247] The modular functional component 16 specifically includes an alarm module 17, a remote control module 18, and a data storage module 19. The alarm module 17 can issue early warning information through audible and visual alarms (dam monitoring center) and SMS push (maintenance personnel); the remote control module 18 can remotely adjust the sensor sampling frequency (e.g., from 1 time / 10min to 1 time / 2min) to ensure data real-time performance; the data storage module 19 is used to store the entire process data in real time.

[0248] Based on this, the visual interactive interface unit 12 in the system is connected and cooperates with the alarm module 17 and the remote control module 18 through the distributed wireless communication module 15 to form the monitoring and decision-making unit of the entire system.

[0249] The following example of an extreme weather (72-hour rainstorm) response implementation plan illustrates the operation of the reservoir dam safety monitoring system in this case.

[0250] This example addresses the scenario of "predicted cumulative rainfall exceeding a threshold (e.g., daily rainfall ≥ 200mm)" and utilizes the collaborative efforts of all system modules to achieve dam stability monitoring, prediction, and emergency dispatch. Figure 3 As shown, the specific steps are as follows:

[0251] 1. Data acquisition stage, which is implemented by the system's intelligent perception layer, mainly involves the system's data acquisition unit 1, multi-parameter detection module 2, high-precision sensing unit 3, three-dimensional seepage detection sensor array 5, and strain sensor 8 based on fiber Bragg grating.

[0252] Among them, the multi-parameter detection module 2 is deployed as follows: a three-dimensional seepage detection sensor array 5 (spacing 10m / unit) is uniformly arranged at different depths inside the dam (such as 0.5m at the dam foundation, 2m in the middle of the dam body, and 1m at the dam top) to synchronously collect three-dimensional distribution data of the seepage field inside the dam body; stress and strain sensors 6 are attached to key parts of the dam foundation rock mass (such as the dam heel and dam toe); and temperature sensors 4 are arranged every 50m on the dam surface to acquire stress and strain (accuracy ±0.1με) and surface temperature gradient (resolution 0.1℃) data in real time.

[0253] High-precision sensor unit 3 is activated: a strain sensor 8 based on a fiber Bragg grating (with electromagnetic interference resistance >60dB and measurement accuracy ±0.1%) is installed in the underwater area of ​​the dam at a depth of 5-10m to monitor the underwater strain of the dam body; a seepage pressure sensor 7 (withstanding water pressure ≥1.2MPa and capable of long-term underwater operation) is deployed at seepage risk points (such as at the seepage prevention curtain) to collect seepage pressure data (resolution 0.1kPa).

[0254] The multi-source data fusion module is linked with 10 stations: it connects to the reservoir area's meteorological and hydrological stations to acquire external data in real time—meteorological data (rainfall, wind speed, wind direction, temperature, sampling frequency 1 time / 10min) and hydrological data (reservoir water level, outflow, water quality, sampling frequency 1 time / 5min).

[0255] 2. Edge preprocessing and data transmission: This stage is mainly completed by the edge computing nodes in the data acquisition unit 1 and the distributed wireless communication module 15.

[0256] Edge computing node processing: The TinyML lightweight model (MobileNetV3) is integrated into each distributed detection unit (corresponding to sensor deployment points) to perform localized preprocessing on the collected raw data.

[0257] Denoising was performed on the seepage pressure and strain data (using wavelet threshold denoising algorithm, the signal-to-noise ratio was improved by 25% after denoising);

[0258] Anomalous data (such as abrupt changes caused by sensor drift) is identified by a self-supervised contrastive learning algorithm, and invalid data is eliminated (the false alarm rate is reduced by 40%).

[0259] Key features (such as seepage rate change rate and strain peak) were extracted, and the data volume was compressed to 20% of the original data.

[0260] Distributed wireless communication module 15 transmission: Adopting a "5G+LoRa dual-link" architecture, each detection unit establishes a relay transmission link through wireless relay nodes (one deployed every 200m, with a signal coverage radius ≥150m) to transmit pre-processed feature data to the central server. Communication latency is <200ms, and bandwidth usage is reduced by 50% compared to traditional full-data transmission. Figure 4 As shown.

[0261] 3. Multi-source fusion and predictive analysis, this stage is completed by the system's intelligent analysis layer, mainly involving data fusion unit 9, multi-source data fusion module 10 and data processing and analysis unit 11.

[0262] Data fusion unit 9 computation: Employing Graph Neural Network (GNN) multimodal fusion technology, a "heterogeneous data graph structure" is constructed—nodes represent sensors / meteorological stations / hydrological stations, and edges represent spatiotemporal correlation weights (e.g., a lag correlation weight of 0.8 for rainfall and water level, and a spatial correlation weight of 0.6 for strain and seepage pressure). This is achieved through the GCN convolutional layer formula:

[0263]

[0264] Achieve cross-modal information transfer and output fused features (comprehensive evaluation accuracy is 25% higher than traditional linear fusion), such as... Figure 5 As shown.

[0265] Data Processing and Analysis Unit 11 Prediction: The LSTM-Transformer hybrid prediction model is enabled (input layer integrates time-series sensor data and external environmental factors). Feature weights are assigned through an attention mechanism (e.g., rainfall weight 0.35, water level weight 0.3, strain weight 0.25). The prediction window is set to 72 hours, and the core indicators are output:

[0266] Dam displacement prediction (error ≤ 0.5 mm, after supplementary optimization ≤ 0.2 mm);

[0267] Crack propagation probability (e.g., a crack propagation probability ≥30% in the dam heel area triggers key attention);

[0268] Prediction of seepage field distribution (predicting the maximum seepage pressure at the anti-seepage curtain after 72 hours).

[0269] Digital twin platform simulation: The fused features and prediction results are input into the digital twin model of the dam to dynamically simulate the stress distribution and seepage path changes of the dam body under rainstorm conditions, and intuitively present the stability status of the dam body (such as the risk area where the stress in the dam toe exceeds the threshold by 15%).

[0270] 4. Emergency decision-making and execution: This stage is mainly completed by the system's decision-making level, and mainly involves the system's visual interactive interface unit 12, alarm module 17, and remote control module 18.

[0271] The reinforcement learning decision engine in the system generates the following strategy: simulating a multi-objective scenario of "flood discharge scheduling - dam stress - water level control", dynamically adjusting the alarm threshold (e.g., lowering the seepage pressure alarm threshold from 100kPa to 85kPa), and outputting the optimal emergency response plan.

[0272] Flood discharge recommendation: Open floodgates No. 2 and No. 3, and control the outflow at 500 m³ / h. 3 / s (to avoid a daily water level rise exceeding 1m);

[0273] Reinforcement priority: Prioritize temporary reinforcement of the dam heel and seepage prevention curtain area (such as sandbag stacking, thickness ≥1.5m);

[0274] Inspection plan: Manual inspection of the dam crest and toe areas will be carried out every 2 hours (in conjunction with GNSS displacement monitoring).

[0275] Modular functional component response:

[0276] Alarm module 17: Issues an "orange alert" via audible and visual alarms (dam monitoring center) and SMS push (maintenance personnel);

[0277] Remote control module 18: Remotely adjusts the sensor sampling frequency (from once / 10min to once / 2min) to ensure data real-time performance;

[0278] Visual interactive interface unit 12: The stress distribution of the dam body simulated by the digital twin is displayed through the virtual reality unit 13, and the location of the reinforcement area is marked through the augmented reality unit 14 (accuracy ±1m). It supports operation and maintenance personnel to view historical data (data on the same period of heavy rain in the past 3 years) and prediction curves.

[0279] 5. Data storage and review: This stage is mainly completed by the data storage module 19 in the system.

[0280] The data storage module 19 stores the entire process data (raw data, preprocessed data, fused features, prediction results, and decision-making schemes) in real time, with a data retention period of ≥5 years, providing data support for subsequent model optimization (such as LSTM-Transformer parameter iteration) and review of similar working conditions.

[0281] The following example illustrates the operation of the reservoir dam safety monitoring system in this case study, using a sudden seepage event (abnormal change in seepage rate) as an example of the implementation plan.

[0282] This example addresses the scenario where "an edge node detects a sudden increase in seepage rate (e.g., from 0.5 L / (m·h) to 5 L / (m·h), exceeding the normal threshold by 10 times)". A rapid response is achieved through "local anomaly localization - global model update - AR-guided intervention". The specific steps are as follows:

[0283] 1. Anomaly detection and localization: This stage is achieved by the system's intelligent perception layer, mainly involving the edge computing node in the data acquisition unit 1, the high-precision sensing unit 3, and the pressure sensor 7.

[0284] Among them, the edge node real-time monitoring: the seepage pressure sensor 7 (numbered S-08) at the downstream seepage curtain of the dam collected seepage pressure data that rose from 80 kPa to 150 kPa within 10 minutes. Simultaneously, the three-dimensional seepage detection sensor array 5 (numbered A-12, located 2m above S-08) detected a sudden increase in seepage rate. The MobileNetV3 model of the edge computing node determined "sudden seepage anomaly" through "abnormal feature matching" (such as seepage rate change rate > 0.5 L / (m·h·min)). The abnormal area was initially located as "300-320m section on the left side of the downstream of the dam, with a depth of 5-8m".

[0285] Data completion and verification: Due to a temporary malfunction of sensor S-07 adjacent to the abnormal area (health level H = 42 < 50), the edge node initiates the IDW data completion algorithm.

[0286]

[0287] Data from three normal sensors (S-06, S-09, and S-10, with spatial distances d of 8m, 10m, and 12m respectively) and historical data from the same period (α = 0.1, reducing historical weight) were used to complete the data for S-07, verifying that the abnormal area had not expanded.

[0288] 2. Global model update and risk assessment. This stage is mainly completed by the intelligent analysis layer in the system, which mainly involves the data processing and analysis unit 11 in the system.

[0289] Based on the federated learning framework, the federated learning collaborative update is completed: each of the 20 distributed detection units trains anomaly detection models locally (based on historical seepage anomaly data), and only uploads the model parameters (not the original data) to the central server, protecting data privacy through differential privacy technology (privacy budget ε = 0.1); after the central server aggregates the parameters, it updates the global anomaly detection model (improving the F1-score to ≥0.92) and distributes it to each edge node to ensure consistent anomaly detection accuracy across the entire dam.

[0290] Data processing and analysis unit 11 evaluation: Combining real-time hydrological data (reservoir water level stable at 145m, no sudden rise) and geological data (abnormal area is sandy gravel strata with high permeability) from multi-source data fusion module 10, the spatiotemporal causal inference algorithm is used:

[0291]

[0292] The calculated causal deviation of "seepage pressure-water level" D is 65% (far exceeding the 15% warning threshold), classifying the risk type as "concentrated leakage caused by localized damage to the anti-seepage curtain," with an estimated leakage volume of approximately 0.8m. 3 / h.

[0293] 3. AR-guided handling and closed-loop management: This stage is mainly implemented by the system's decision-making layer, and mainly involves the visual interactive interface unit 12 and the remote control module 18.

[0294] AR system path guidance: The augmented reality unit 14 of the visualization interactive interface unit 12 generates an "abnormal area navigation path" - maintenance personnel can view the real view of the dam through AR glasses, and overlay the location of the seepage point (accuracy ±0.5m), the distribution of surrounding sensors, and safety passages; at the same time, "repair process guidance" is displayed (such as "step 1: clean up the debris around the seepage point → step 2: inject cement slurry (water-cement ratio 1:1.5) → step 3: lay geomembrane").

[0295] Remote control and effect verification: The sampling frequency (1 time / 1min) of the pressure sensor 7 and stress strain sensor 6 around the abnormal area is remotely adjusted through the remote control module 18 to monitor the repair effect in real time; 2 hours after the repair, the pressure drops to 90kPa and the seepage rate recovers to 0.6L / (m·h). The data processing and analysis unit 11 determines that "risk is eliminated", the alarm module 17 terminates the warning, and the system returns to normal monitoring mode.

[0296] As can be seen from the above, the solution presented in this example has the following advantages over existing technologies:

[0297] 1. Data collection: More comprehensive and accurate, adaptable to complex environments.

[0298] Existing technologies rely on single sensors (such as independent flowmeters or strain gauges) to collect data, resulting in limited coverage, susceptibility to electromagnetic interference, and low data accuracy. This example solution utilizes a distributed high-precision sensor array combined with anti-interference design to achieve comprehensive and accurate data acquisition.

[0299] Deploying a three-dimensional seepage detection sensor array (uniformly distributed at different depths and locations within the dam) can capture the three-dimensional distribution of the seepage field inside the dam, avoiding "local data loss";

[0300] Employing a strain sensor based on a fiber Bragg grating (with electromagnetic interference resistance >60dB and measurement accuracy ±0.1%), it can still stably output high-precision strain data in harsh electromagnetic and underwater environments.

[0301] The multi-parameter detection module (seepage, stress, temperature) collects data simultaneously, covering multiple dimensions of data from "dam body to dam foundation to external environment", providing a complete data foundation for subsequent analysis.

[0302] 2. Data fusion: Breaking through the limitations of linearity, enabling spatiotemporal correlation analysis.

[0303] Existing technologies employ "linear weighted fusion" of multi-source data, which fails to capture the spatiotemporal coupling relationships between data (such as the lag correlation between rainfall and dam seepage). This example solution utilizes graph neural network (GNN) multimodal fusion technology to construct a graph structure with "sensor / environmental factors as nodes and spatiotemporal correlations as edges," enabling cross-modal information transfer.

[0304] It can integrate sensor data, meteorological and hydrological data (rainfall, water level, wind speed), and geological model data to explore the dynamic coupling relationship between the dam structure and the external environment;

[0305] The overall assessment accuracy is 25% higher than that of existing linear fusion methods, solving the problem of "missing to identify associated risks when looking at only a single data point" (such as the correlation analysis between rainstorms and sudden increases in seepage pressure).

[0306] 3. Predictive analysis: higher accuracy, longer lead time, and coverage of extreme working conditions.

[0307] Existing technologies employing traditional machine learning (such as regression analysis) have weak modeling capabilities for complex nonlinear relationships, low prediction accuracy (dam displacement prediction error ≥2mm), and cannot cover extreme working conditions. This example solution achieves high-precision long-term prediction through deep learning model optimization.

[0308] By adopting the LSTM-Transformer hybrid model and incorporating an attention mechanism (automatically assigning feature weights), and combining historical and real-time data, the dam displacement prediction error is ≤0.5mm (which can be reduced to ≤0.2mm after optimization).

[0309] It supports multi-step prediction for ≥72 hours, and the training data covers extreme conditions such as "once-in-a-century floods". It can predict the risk of dam deformation and crack expansion under extreme conditions in advance, avoiding the passivity of "post-event response".

[0310] 4. Communication transmission: more real-time and efficient, reducing bandwidth pressure.

[0311] Existing technologies transmit raw data in full via wired connections, resulting in high bandwidth consumption and significant latency (failing to meet rapid response requirements). This example solution significantly improves real-time performance and transmission efficiency through edge computing and distributed wireless communication.

[0312] Lightweight AI models (such as MobileNetV3) are deployed on edge computing nodes, and the raw data is preprocessed locally (denoising, anomaly detection). Only key feature data is transmitted, the data volume is compressed to 20% of the original data, and bandwidth usage is reduced by 50%.

[0313] It adopts a 5G / LoRa distributed wireless communication architecture, coupled with wireless relay nodes to expand the signal coverage, with a communication latency of <200ms and a total process time of <1s from "data acquisition to policy generation", meeting the rapid response requirements of emergency scenarios.

[0314] 5. Privacy protection and collaborative monitoring: Balancing data security with model generalization.

[0315] Existing technologies require the transmission of raw data during multi-department collaborative monitoring, posing a risk of data privacy breaches. This example solution introduces a federated learning framework to resolve the conflict between collaborative monitoring and privacy protection.

[0316] Each detection unit trains its model locally and only uploads model parameters (not raw data) to the central server to avoid leakage of sensitive data (such as dam structure parameters and reservoir hydrological data).

[0317] The central server aggregates parameters from multiple units, ensuring the model's generalization ability and enabling "multi-department collaboration without data interoperability," making it suitable for dam monitoring scenarios involving cross-regional and multi-entity management.

[0318] 6. Anomaly Detection and Decision Making: Smarter, from “passive early warning” to “proactive decision making”.

[0319] Existing technologies rely on fixed threshold alarms, resulting in high false alarm rates and only providing passive alerts without supporting emergency response strategies. This example solution utilizes adaptive algorithms and reinforcement learning to achieve intelligent anomaly detection and proactive decision-making.

[0320] The self-supervised contrastive learning algorithm is used to identify data anomalies (such as sensor drift and sudden seepage) without supervision, eliminating the need for manual labeling and reducing the false alarm rate by 40%.

[0321] The integrated reinforcement learning decision engine can simulate emergency scenarios such as sudden changes in seepage and stress exceeding limits, dynamically adjust alarm thresholds (such as lowering the seepage pressure alarm threshold during the flood season), and generate optimal response strategies (such as flood discharge flow suggestions and reinforcement priority ranking), realizing a closed loop from "detecting anomalies" to "solving anomalies".

[0322] 7. Visualized Interaction: More intuitive, lowering the barrier to entry for operation and maintenance.

[0323] Existing technologies lack dedicated visualization tools, presenting data in tables and simple curves, making it difficult for maintenance personnel to intuitively understand the dam's safety status. This example solution enhances operational convenience through VR / AR visualization and interactive design.

[0324] The visual interactive interface unit integrates a virtual reality (VR) unit, which can dynamically display abstract data such as dam stress distribution and seepage path;

[0325] Augmented Reality (AR) units support "real-scene overlay," such as marking the location of leaks and displaying repair process guidelines in a real-world dam scene, helping maintenance personnel quickly locate problems and take action, thus lowering the barrier to professional operation.

[0326] 8. Economy and Reliability: Reduce costs and improve incident response efficiency.

[0327] Existing technologies suffer from high maintenance costs and delayed incident response due to their reliance on full-volume transmission and manual inspection. This example solution achieves cost reduction and efficiency improvement through technical optimization.

[0328] Edge computing reduces data transmission volume, wireless communication replaces wired cabling, and maintenance costs are reduced by 30%.

[0329] Early warning (such as predicting risks 72 hours in advance of extreme weather) and rapid decision-making (generating strategies within 5 minutes of sudden seepage) improve accident response efficiency by 90%, avoiding economic losses and safety risks caused by delayed handling.

[0330] In summary, the solution of this invention, through multi-stage technological innovation, achieves significant advantages in data accuracy, predictive capability, real-time performance, and intelligent decision-making. It not only addresses the core pain points of existing technologies but also constructs a complete intelligent monitoring system encompassing "perception-fusion-analysis-decision-interaction," forming a technological barrier and demonstrating extremely high inventiveness.

[0331] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of this invention is defined by the appended claims and their equivalents.

Claims

1. A reservoir dam safety monitoring system, characterized by, The system comprises: A distributed sensing unit comprising a plurality of sensing modules deployed on the dam body of a reservoir dam to be monitored, and an edge computing node module for processing data collected by the plurality of sensing modules, wherein an edge robust mechanism is configured in the edge computing node module to bind the sensor health status and data validity in real time; A data fusion unit connected with the distributed sensing unit, performing hierarchical division processing on the received sensing data to form multi-scale data, and constructing a corresponding scale graph structure for each level of scale data, and then performing fusion through a multi-scale graph convolution fusion model; A data processing and analysis unit connected with the data fusion unit, and configured with a spatio-temporal causal inference early warning model and a dynamic spatio-temporal attention prediction model; The spatio-temporal causal inference early warning model generates early warning information when the data generated by the data fusion unit is within the threshold but the causal logic is abnormal by constructing a causal relationship network of dam safety; The dynamic spatio-temporal attention prediction model calculates the risk weight of the fusion data generated by the data fusion unit by partitioning, and adjusts the weight by time period, and finally performs hybrid prediction calculation according to the weighted features and the identified working conditions; A monitoring decision unit interacting with the data processing and analysis unit, constructing a corresponding scene model based on meta-learning for a target reservoir dam, performing twin synchronization simulation based on the warning and prediction data analyzed by the data processing and analysis unit, calculating the scene similarity, fine-tuning the model according to the similarity value, and finally generating a decision scheme according to the determined scene model.

2. The reservoir dam safety monitoring system of claim 1, wherein, The edge computing node module is configured with a fault diagnosis submodule, a data completion submodule, and a dynamic path switching submodule, The fault diagnosis submodule analyzes the data fluctuation characteristics by inputting sensing data, and calculates the sensor health value based on the characteristic data obtained by analysis; The data completion submodule interacts with the fault diagnosis submodule, and can complete the missing data of the sensing data by inverse distance weighting method when the fault diagnosis submodule detects the failure of the corresponding sensing module; The dynamic path switching submodule automatically switches to the backup sensing module based on the preset sensing path for the fault sensing module determined by the fault diagnosis submodule.

3. The reservoir dam safety monitoring system of claim 1, wherein, The multi-scale graph convolution fusion model comprises a scale layering module and a scale attention module, The scale layering module is used to perform multi-level layering on the transmitted data, and generate corresponding multi-level scale graph structures and assign initial weights; The scale attention module is used to dynamically adjust the scale weights by calling the sigmoid function for the multi-level scale graph structures generated by the scale layering module.

4. The reservoir dam safety monitoring system of claim 1, wherein, The spatio-temporal causal inference early warning model is configured with a causal relationship network module, a real-time causal bias monitoring module, and a risk trace positioning module, The causal relationship network module is set to mine strong causal relationships between key parameters of the dam through a causal discovery algorithm to form a core causal chain; The real-time causal bias monitoring module is configured to interact with the causal relationship network module, calculate the bias degree of each causal link generated by the causal relationship network module, and monitor each calculated link bias degree. When the bias degree exceeds a threshold, a warning is triggered. The risk traceability positioning module is configured to interact with the real-time causal bias monitoring module. After the real-time causal bias monitoring module triggers a warning, the risk source is located through a causal reverse algorithm for the causal link with a bias degree exceeding the threshold.

5. The reservoir dam safety monitoring system of claim 1, wherein, The dynamic spatio-temporal attention prediction model is configured with a spatial dynamic attention module, a temporal dynamic attention module, and a hybrid prediction output module. The spatial dynamic attention module is configured to divide the dam into N monitoring units. The risk contribution degree of each unit is calculated to dynamically adjust the attention weight. The temporal dynamic attention module is configured to divide the time series into sensitive time periods, assign a sensitivity coefficient to each sensitive time period, and adjust the weight through a time period sensitivity function. The hybrid prediction output module selects the optimal prediction sub-model according to the real-time working condition by introducing real-time working condition data to generate a deformation prediction value.

6. The reservoir dam safety monitoring system of claim 1, wherein, In the meta-training phase, the meta-reinforcement learning dynamic decision engine constructs a dam burst scenario library and trains a basic decision model using the MAML algorithm.

7. The reservoir dam safety monitoring system of claim 6, wherein, In the real-time decision-making phase, the meta-reinforcement learning dynamic decision engine first performs digital twin synchronization to input real-time sensor data and external environment data into the dam digital twin model to simulate the stress state of the dam body under the current scenario. Then, scene matching and fine-tuning are performed to compare the similarity of the current scenario and the scenario library. If the similarity is less than the corresponding threshold, the meta-learning algorithm is called to fine-tune the decision model based on the virtual samples simulated by the digital twin. Finally, multi-objective decision output is generated to generate a decision scheme for the new scenario.

Images